Embeddings Visualizer

A tool for visualising token embedding spaces.

Each axis is specified with a positive and negative token. A word list is then projected onto this axis.

See a more detailed explanation here.

Frontend uses D3.js.

Backend uses Python, flask, and gensim.

Deployed on a DigitalOcean VPS.

Interactive Map of AI Safety Ecosystem

I was comissioned to create an interactive map of the ecosystem of organisations related to AI safety.

Users can zoom and pan across the map, hover over items for more information, and click on items to see the respective website.

Created with D3.js.

The data is pulled from a Google Sheet so that anyone can suggest updates or corrections.

Animated Income Visualisations

I was commissioned to create animated visualisations of US income over time.

The data preparation and prototyping was done ith Python, Pandas, and pyplot.

The polished visualisations were created with html, and D3.js.

A web interface in the repo allows further customization of the animations.

Litmaps

Litmaps is an early stage startup which builds tools for visualising and discovering scientific papers.

I coded the visualisation logic currently in production.

This involved implementing efficient javascript algorithms for visualisation use-cases and setting up any necessary backend using FastAPI and Python.

More details can be found here.

Animation: Mechanistic Interpretability Overview

I animated a snippet from the podcast AXRP to provide an overview of the motivation and content of mechanistic interpretability.

The podcast is a conversation between researcher Neel Nanda and host Daniel Filan.

This was used in a training course and has 5.9k views.

Animation: Modular Arithmetic Mehcanistic Interpretability

I animated a snippet from the podcast AXRP which illustrates how transformers do modular arithmetic, which was an output of some of Neel Nanda's mechanistic interpretability research.

The podcast is a conversation between researcher Neel Nanda and host Daniel Filan.

Detecting Concept Drift in Medical Triage

2020 ACM SIGIR Conference on Research and Development in Information Retrieval

I developed an algorithm for detecting concept drift (a species of data shift) via model miscalibration.

The paper uses machine learning medical triage models as a motivating example.

See the below section on my master's thesis for more context.

Concept Drift in Medical Referrals Triage

2020 M.Sc. thesis

My M.Sc. was funded by a scholarship form the Presicion Driven Health Partnership. I worked on detecting when a medical machine learning model has become obsolete and requires retraining.

I won the annual "best M.Sc." prize from the Auckland University computer science department. One of my examiners said it was "the best master's thesis they had seen".

Feature Importance for Biomedical Named Entity Recognition

2019 Australasian Joint Conference on Artificial Intelligence

I did this research during a research intership at Orion Health.

The paper surveys features which have been used in biomedical natural language processing, and evaluates each feature's utility for a deep learning approach to biomedical named entity recognition.

Predicting Air Quality from Low-Cost Sensor Measurements

2018 Australasian Conference on Data Mining

I worked with the National Institute of Water and Atmospheric Research (NIWA) on modelling air quality using low cost sensors.

Low-cost sensors pose a number of challenges: they are unreliable, noisy, and may require ongoing calibration. This paper explores these challenges, and evaluates several approaches to spatiotemporal modelling of air quality using the sensors.